An Interpretable Data-Driven Polynomial Regression Model for Early-Stage Prediction of Reinforced Concrete Floor Cost in Civil Buildings

Abstract

Accurate early-stage cost estimation is essential for value engineering and preliminary decision-making in civil building projects. However, conventional estimating approaches often rely on detailed design information or simplified linear assumptions, which limit their usefulness at the conceptual stage. This study develops an interpretable data-driven polynomial regression framework for predicting reinforced concrete floor cost in civil buildings. A dataset of 4477 samples derived from RSMeans floor-system assemblies was used, including slab type, tributary area, superimposed load, formwork price, and concrete price as predictors. Four candidate polynomial regression models with degrees 1 to 4 were constructed and compared using a 70:30 training-testing split and 1000 Monte Carlo simulations based on the coefficient of determination, root mean square error, and mean absolute error. The results show that predictive performance improves consistently as the polynomial degree increases. Although the quartic model achieved the highest raw predictive accuracy, the cubic model was selected as the final model because it provided the most appropriate balance between predictive performance and interpretability. The selected model achieved R²=0.965, RMSE = 8.764 USD/m², and MAE = 6.163 USD/m² on the training dataset, while the corresponding testing results were R²=0.965, RMSE = 8.567 USD/m², and MAE = 6.058 USD/m². Feature-importance analysis, centered partial dependence plots, and contour plots further showed that formwork and concrete are the dominant cost drivers and that the response surface is strongly nonlinear and interaction-driven. The proposed framework provides a transparent and practically usable tool for early-stage floor cost prediction and engineering decision support.